Parametric study on the accuracy of full-field reconstruction from sparse measurements using autoencoders

Abstract

Full-field data provides a comprehensive understanding of the behavior of a system or structure, which is particularly crucial when identifying local damages. These damage may exhibit complex and subtle effects that could be overlooked with sparse measurements. Recent advancements in machine learning, such as autoencoders (AE), have enabled the reconstruction of full-field data using sparse measurements. However, a study assessing the accuracy of AE in reconstructing full-field data concerning measurement locations, data sparsity, and noise density is still lacking in the context of nondestructive evaluation (NDE). To address these gaps, this study adopts a parametric approach to evaluate the effectiveness of an LSTM-based AE model in terms of measurement locations, data sparsity, and noise density. The two sets of data (i.e., configuration #1 and #2) were generated using a finite element method for a 2D metallic plate cooling. The configuration #1 data were then used to train the LSTM-based AE and the model's full field reconstruction performance was validated on sparse measurements of configuration #2 using the average reconstruction error (ARE) as a testing parameter. The result shows, there was no significant impact of different measurement locations on ARE. Whereas ARE increased with increase in data sparsity and noise density. This research presents a parametric study with potential applications in full-field reconstruction, not limited to thermal data. It can be extended to other applications, such as strain, displacement, and velocity, in scenarios where the targeted system undergoes temporal evolution.